CD Laboratory for Steel Industry Signal Processing and Machine Learning

The aim of this CD Laboratory is to advance and improve the digitisation of processes in the steel industry. Many industrial plants are already monitored by sensors of all kinds, but the sensor data collected often contains significant noise due to the harsh environment. Research in this CD Laboratory aims to improve the usability of such noisy signal data. The research results should be widely applicable to a variety of plants and processes by identifying recurring signal characteristics, e.g. periodic interference signals, across many measuring points and searching for broadly applicable solutions.

The manufacturing processes for steel products, such as steel strips in various thicknesses, railway rails and foundry products, require many different industrial plants such as blast furnaces, foundries, rolling mills, etc. In addition to mechanical components, these industrial plants contain many different electronic systems for controlling and monitoring the machines and for fault detection. These electronic systems consist of sensors and cameras, as well as computers that process all available information using software and algorithms. However, due to the harsh industrial environment, the sensor signals are often affected by strong noise, which reduces the amount of useful information. In other cases, unwanted signal components occur that require special evaluation before the sensor signals can be further processed. There are cases where the lack of accurate signal models prevents the extraction of detailed information from the sensor signals. Even with accurate signal models, the extraction of the desired information is sometimes hindered by the lack of computationally feasible and accurate algorithms.

This CD Laboratory therefore has three overarching goals:

1. The algorithms that process the sensor signals must be able to cope with the low signal-to-noise ratio (SNR), unwanted signal components and other undesirable effects. The design of algorithms that deliver accurate results under these conditions is an open research topic. While classical model-based approaches are being investigated, machine learning-based methods or hybrid approaches are currently the dominant trend. One goal of this CD Laboratory is to advance research on model-based algorithms and machine learning-based methods and their combinations, with a focus on low SNR scenarios.

2. In a large industrial plant, numerous different sensors continuously measure various signals. Hundreds of computers execute a variety of algorithms to process these sensor signals. Despite this diversity, patterns can be identified, e.g., a kind of periodicity in the signal properties or an unwanted background signal, which may also be periodic. Finally, some algorithms may be the same, e.g. estimation algorithms that estimate the fundamental frequency of a periodic signal. Therefore, improvements to an algorithm or its signal processing chain can be used to improve the accuracy of signal processing at several other locations within the plant. One goal of this CD Laboratory is therefore to provide broadly applicable algorithms, insights and performance limits.

3. The transition from purely mechanical industrial plants in the past to highly digitised and computerised industrial plants is far from complete. Even today, there is still a great need to monitor processes, measure properties and replace existing measurement systems and their algorithms with better ones. One goal of this CD Laboratory is therefore to drive forward the ongoing digitisation.